OpenAI Launches Dot: An Always-On Personal Agent for Work and Home

OpenAI has introduced Dot, a personal Agent designed to keep track of your goals around the clock. It is more than a chat tool that waits for you to ask: once you give it a goal and connect the relevant apps, it can watch for developments while you are away from the chat, identify work that needs doing, and bring back results or actions that need your approval. In OpenAI's demonstration, Dot follows both work projects and family plans.

Its main capabilities fall into four groups:

• Spot follow-up work: Watch for changes in authorized apps, such as a recurring customer complaint or a changed plan.
• Prepare the work itself: Research, use a browser, write code, and run tests in its own cloud computer, rather than only offering advice.
• Track several goals at once: Retain the state of different projects and continue the same work across ChatGPT, Slack, or Teams.
• Leave consequential decisions to you: Sending messages, making purchases, and permanently deleting data require approval. You can inspect its progress and pause it.

The details matter: how does Dot discover a task, how far can it take the work, and what should you know before granting access?

The biggest difference from ChatGPT: who notices the work first

With an ordinary ChatGPT conversation, you generally have to notice a problem yourself, then supply the materials and instructions. Once Dot has an ongoing assignment, it can notice changes in authorized apps. If customers report the same bug again, for example, it can investigate, prepare a fix, and submit a PR for your review. You need not open a fresh conversation and explain the project from scratch every time.

OpenAI describes an early tester who forgot to invoice a publisher. Dot noticed the omission and prepared an invoice, then sent it only after the tester approved. Dot handled the follow-up and preparation; the final decision stayed with the person.

OpenAI calls the background search for leads “proactive research.” This stage uses read-only tools: it can inspect information in connected apps, but cannot send messages, modify app content, or control your computer. Acting on what it finds is a separate step, governed by app permissions and action rules.

To keep working over time, Dot needs to retain at least three kinds of context:

• Goal state: What the project is meant to accomplish, what is blocking it, and what comes next.
• Personal standards: Your preferred way of communicating and judging whether a result is good enough.
• Work state: What changed in files, apps, and tasks; what has been done; and what still needs your approval.

The promotional film cuts between work and home: Dot updates a website, board materials, and code, while also dealing with a wedding cake, after-school activities, and a family calendar. The user still changes direction, reviews results, and approves important actions.

How Dot keeps a task moving

Dot is powered by GPT-6 Astra. It has its own cloud computer and browser, and can write code and run tests when needed. You can open that computer to inspect its progress. Access to your local machine is off by default; Dot can use files and a browser on your own computer only after you explicitly authorize that in the desktop app.

It can also connect to apps. OpenAI says its plugin ecosystem covers more than 4,000 apps, but you still choose which ones to connect and what permissions to grant. These connections can give Dot background on a project. If you continue the work in ChatGPT, Slack, or Teams, you do not have to explain everything again. Dot can retain progress on several projects and pick up where it left off when you return.

When one thing changes, related materials must change too

OpenAI shows five kinds of work. They share one problem: something changes, related materials need updating, and people often miss one of them.

When developers receive recurring customer feedback, Dot can define the issue, fix and test it, then deliver a PR and demo video for review. If the scope of a product launch changes at the last minute, it can identify which copy, assets, and documents need revisions and prepare drafts.

If research data produces an unexpected result, it can rerun the analysis, update charts and explanations, and flag what researchers need to check. If a sales prospect changes its requirements, it can check documents, test an integration, and revise the proposal and open questions. If a creator changes their editing notes, it can carry those changes into show notes and social-post drafts. When you change one thing, Dot can help find the other things that need changing too.

These scenarios clarify what OpenAI wants Dot to take on. They do not yet tell us how reliably it performs when tasks multiply or circumstances change, or how much time it takes to clean up mistakes.

Can the same work continue in another channel?

Dot can report progress or ask you to make a decision. In OpenAI's SMS demo, it uses a calendar and a menu to propose a dinner plan, calculates the price, tax, and tip, and explicitly says it has not ordered anything yet. It waits for the user's confirmation. The task starts with a calendar and reaches a decision by text message, without requiring the user to restate the context.

Availability differs by channel. You can create Dot in ChatGPT on a computer, then use it in the mobile app after setup; the mobile web is not supported yet. SMS is a limited test for US Pro users, passes through a third-party provider, and is unavailable in Business and Enterprise workspaces. Dot cannot proactively call you yet. The OpenAI Help Center lists these conditions.

What can it access, and when must it ask?

Dot can encounter information in apps and may operate a browser or write code. OpenAI describes several layers of control:

OpenAI's safety video gives examples: a new purchase, permanent deletion, or message to someone else requires approval; changing a password must be done by the user. The activity view lets you inspect past and ongoing work and pause Dot.

These controls also have practical limits. Disconnecting an app does not delete information Dot has already obtained from it. The Help Center says that removing Dot's saved memory requires resetting the entire Dot, which also deletes its conversations and scheduled tasks. Before connecting email, files, or customer systems, consider what Dot will be allowed to read. OpenAI also advises people to check important outputs.

The day before Dot's announcement, OpenAI paused the release of GPT-6.1 Astra. CBS reported safety concerns involving scope and authorization. Dot currently uses GPT-6 Astra, not the paused model. These are separate developments, though both underline why an Agent's permission boundaries need to work in practice.

Model training depends on the plan and the data path

Content in Business, Enterprise, and Edu workspaces is not used to improve models by default. Users on personal plans can control whether Dot conversations and work content may be used for training.

Proactive-research content and Dot's own notes are not used directly for training. If that information later enters a conversation or task eligible for training, it may still be used according to your data settings. Dot also reads existing ChatGPT memories and forms its own memory from connected apps. Simply disconnecting an app does not withdraw those saved details.

OpenAI is also piloting a “dedicated Dot” for enterprises: it works on behalf of a defined role with its own identity, credentials, and system access. For now, OpenAI is working with a small number of companies to define responsibilities, tools, and human review. Integration with Microsoft Agent 365 governance controls is still in progress.

Can you use Dot now? It depends on market, plan, and limits

For now, Dot must be created in ChatGPT on a computer, and access is rolling out in stages. The OpenAI Help Center says the initial Pro rollout excludes the European Economic Area, Switzerland, and the UK; Business Premium covers supported ChatGPT regions; and the Enterprise beta is off by default until an administrator enables it. Pro and Business Premium each include the first Dot.

Usage limits are easy to misread. The Help Center says eligible users' Dot usage does not count against plan limits during the first month, with longer-term rules to be announced later. The launch post also says Codex and ChatGPT work tasks initiated or managed by Dot still count against those products' respective limits. The two statements refer to different kinds of usage. “One Dot included” does not mean every task is unlimited, and a free first month does not establish the long-term cost.

The promotional film runs smoothly, but an Axios report from the event noted a noticeable delay during one voice call to Dot. A single delay says little about long-term performance. Between that film and the product you can actually use lie both the phased rollout and the reality of a live demonstration.

Ordinary users start with one Dot. Teams of multiple Dots are a future direction OpenAI has described, not an available feature.

Is it worth handing work to Dot?

A missed invoice, a recurring bug report, a chart that was not updated after the data changed: none is especially hard on its own. The hard part is having someone remember every follow-up. Dot's appeal lies there. It can watch for changes and put the next piece of work in front of you, or prepare a version for you to review.

Whether it really saves time depends on whether it misses things or makes incorrect changes, and how much effort you spend checking its results. For now, give it work with clear boundaries and outcomes you can verify easily. Before connecting more apps or letting it touch commitments to other people, examine its actual performance and permission record. Four questions are enough to start: What can it read? What can it change? Which actions require my approval? Where can I trace a mistake?